Case Study: OPPO achieves up to 360% faster AI training with CNCF CubeFS

How CubeFS Accelerates AI training 3x in OPPO Hybrid Cloud Platform

OPPO, a leading consumer electronics company, faced significant storage performance, cost, and security challenges when building a hybrid GPU cloud for its AI training platform. Network latency between its private cloud and the public cloud caused sharp decreases in storage performance and low GPU utilization when tasks were offloaded. To tackle these issues, OPPO turned to the Cloud Native Computing Foundation's CubeFS, a cloud-native distributed storage system.

By implementing the CubeFS client with its caching function, OPPO cached training data directly on the GPU compute nodes. This solution, provided by the Cloud Native Computing Foundation, reduced metadata latency to microseconds and drastically accelerated AI training. The measurable impact included a performance improvement of up to 360% with the RESNET18 model and 130% with AlexNet, while also providing a unified, secure storage layer across their hybrid cloud that eliminated the need for separate dataset copies.

OPPO